Vehicle driving intention prediction method and device, vehicle, medium and program product

Abstract

The application relates to a vehicle driving intention prediction method and device, a vehicle, a medium and a program product, which are applied to a vehicle and include the following steps: taking a target vehicle as a center to screen vehicles within a target distance as graph nodes, and extracting multi-dimensional motion features of each vehicle to form a node attribute matrix; all vehicle pairs are traversed, an edge is constructed if the distance between the vehicle pairs is less than a preset distance, an edge weight corresponding to the edge is calculated, and an initial adjacency matrix is generated; based on the node attribute matrix and the initial adjacency matrix, a multi-stage progressive graph self-enhancement processing is performed, a graph convolution network (GCN) and adaptive parameters of a corresponding layer are loaded at each stage, and a processed target vehicle feature is obtained; and the processed target vehicle feature is input into an intention classifier to obtain a driving intention output by the intention classifier. The application performs the multi-stage progressive graph self-enhancement processing, layer by layer extracts full-scene interaction semantics from a basis to complexity, greatly improves the distinguishing degree and robustness of the feature, and improves the prediction accuracy.

Description

Technical Field

[0001] This application relates to the field of vehicle technology, specifically to a method and device for predicting vehicle driving intentions, a vehicle, a medium, and a program product. Background Technology

[0002] Vehicle driving intent prediction is one of the core functions of autonomous driving systems. It aims to predict the driving behavior that the vehicle will perform in advance by analyzing the motion state of the target vehicle and the surrounding traffic participants, thus providing a basis for autonomous driving decisions.

[0003] In recent years, Graph Neural Networks (GNNs) have been widely used in driving intention prediction tasks due to their ability to effectively model the interaction relationships between vehicles. However, related GNN-based driving intention prediction methods still have the following limitations: as the number of GNN layers increases, node features tend to become homogenized, making it difficult for the model to distinguish vehicle features with different driving intentions.

[0004] It is evident that the relevant technologies suffer from technical defects in predicting driving intentions, such as insufficient interaction modeling and oversmoothing during training, which limits prediction accuracy. Summary of the Invention

[0005] One objective of this application is to provide a vehicle driving intention prediction method to solve the problems of insufficient interactive modeling and oversmoothing during training in related technologies, which lead to limited prediction accuracy; another objective is to provide a vehicle driving intention prediction device; a third objective is to provide a vehicle; a fourth objective is to provide a computer-readable storage medium; and a fifth objective is to provide a computer program product.

[0006] To achieve the above objectives, this application provides a method for predicting vehicle driving intentions, applied in a vehicle, and the technical solution adopted is as follows: Vehicles within the target distance are selected as graph nodes centered on the target vehicle. Multi-dimensional motion features are extracted from each vehicle to form a node attribute matrix. All vehicle pairs are traversed. If the distance between vehicle pairs is less than a preset distance, an edge is constructed, and the edge weight is calculated to generate an initial adjacency matrix. Based on the node attribute matrix and the initial adjacency matrix, a multi-stage progressive graph self-enhancement process is performed. At each stage, a graph convolutional network (GCN) with the corresponding number of layers and adaptation parameters are loaded to obtain the processed target vehicle features. The processed target vehicle features are then input into an intent classifier to obtain the driving intent output by the intent classifier.

[0007] Based on the above technical means, multi-stage progressive graph self-enhancement processing can be performed to progressively extract interactive semantics from basic to complex scenarios, greatly improving the discriminativeness and robustness of features and enhancing prediction accuracy.

[0008] Furthermore, based on the node attribute matrix and the initial adjacency matrix, a multi-stage progressive graph self-enhancement process is performed. At each stage, a graph convolutional network (GCN) of the corresponding number of layers and adaptation parameters are loaded to obtain the processed target vehicle features, including: Based on the node attribute matrix and the initial adjacency matrix, the first stage loads two layers of GCN and the first stage adaptation parameters for processing, and outputs first-order interaction features and car-following recognition results. Based on the first-order interaction features and the updated adjacency matrix obtained in the first stage, the second stage loads the 3-layer GCN and the adaptation parameters of the second stage for processing, and outputs the second-order interaction features and lane change recognition results. Based on the second-order interaction features and the second-stage updated adjacency matrix, the third stage loads the 4-layer GCN and the third-stage adaptation parameters for processing, outputting high-order interaction features and full-scene intent recognition results; among which, the processed target vehicle features include high-order interaction features.

[0009] Based on the aforementioned technical means, basic car-following interaction features are quickly extracted, and 32-dimensional first-order interaction features and binary classification car-following recognition results are output, avoiding the premature oversmoothing problem in deep networks. While preserving car-following semantics, motion information of vehicles in adjacent lanes is fused to output 32-dimensional second-order interaction features and three-class classification lane-change recognition results, balancing convergence speed and feature stability. Covering multi-vehicle game information across the entire scenario, 32-dimensional high-order interaction features and 6-class classification full-scene intent recognition results are output. Finally, the line vector corresponding to the target vehicle is extracted as the processed target vehicle feature, achieving high-precision intent prediction in complex scenarios.

[0010] Furthermore, based on the node attribute matrix and the initial adjacency matrix, a two-layer GCN and the adaptation parameters of the first stage are loaded and processed in the first stage to output first-order interaction features and car-following recognition results, including: Based on the node attribute matrix, two layers of GCN are loaded in the first stage. The first layer of GCN maps the multidimensional motion features to 64-dimensional motion features, and the second layer of GCN maps the 64-dimensional motion features to 32-dimensional motion features. Among them, the first-order interaction features are 32-dimensional motion features. In the first stage, the homogeneity enhancement threshold, homogeneity suppression threshold, enhancement coefficient, and suppression coefficient of the first stage are loaded, and the initial adjacency matrix is ​​dynamically updated to obtain the updated adjacency matrix. In the first stage, the dynamic feature calibration coefficients of the first stage are loaded, and the node features of the first stage are calibrated to the centroids of the two categories of car-following and non-car-following to obtain the car-following recognition result.

[0011] Based on the above technical means, only the car-following interaction information of directly adjacent vehicles is focused on to avoid introducing irrelevant noise and quickly complete the extraction of basic interaction semantics; the car-following related edge connections are strengthened and irrelevant interference edges are weakened, so that the graph structure is more in line with the real interaction logic of the car-following scenario; the large step size allows the scattered features to quickly converge to the two types of centers, which greatly improves the convergence speed and recognition accuracy of car-following binary classification.

[0012] Furthermore, based on the first-order interaction features and the updated adjacency matrix obtained in the first stage, the second stage loads a 3-layer GCN and the adaptation parameters from the second stage for processing, outputting second-order interaction features and lane-change recognition results, including: In the second stage, a 3-layer GCN is loaded. By reusing the weights of the first and second layers of GCN, the 32-dimensional motion features are processed to obtain the reconstructed features under the updated adjacency matrix. The reconstructed features are aggregated through the third-layer GCN to obtain 32-dimensional second-order interaction features, where the second-order interaction features are 32-dimensional second-order interaction features. In the second stage, the homogeneity enhancement threshold, homogeneity suppression threshold, enhancement coefficient, and suppression coefficient of the second stage are loaded, and the updated adjacency matrix is ​​dynamically updated to obtain the adjacency matrix after the second update. In the second stage, the dynamic feature calibration coefficients of the second stage are loaded to calibrate the centroids of the second stage node features to three categories: lane keeping, left lane changing, and right lane changing, so as to obtain the lane change recognition results.

[0013] Based on the above technical means, the learned car-following basic interaction capabilities are fully preserved, avoiding the forgetting problem caused by training a 3-layer GCN from scratch, and quickly adapting to the graph structure of lane-changing scenarios; the motion information of vehicles in adjacent lanes is fused to encode the multi-vehicle game relationship in lane-changing scenarios, providing high-order semantic support for subsequent lane-changing recognition; the lateral vehicle connections related to lane changing are strengthened, further optimizing the graph topology structure and adapting to the interaction distribution of lane-changing scenarios; the convergence speed and feature stability are balanced to avoid feature oscillations caused by large step sizes, significantly improving the recognition accuracy of lane-changing three-class classification.

[0014] Furthermore, based on the second-order interaction features and the second-stage updated adjacency matrix, the third stage loads a 4-layer GCN and the third-stage adaptation parameters for processing, outputting high-order interaction features and full-scene intent recognition results, including: In the third stage, the homogeneity enhancement threshold, homogeneity suppression threshold, enhancement coefficient, and suppression coefficient of the third stage are loaded, and the adjacency matrix after the second update is dynamically updated to obtain the adjacency matrix after the third update. The adjacency matrix after three updates is normalized to obtain the normalized adjacency matrix; In the third stage, four layers of GCN are loaded. By reusing the weights of the first, second and third layers of GCN, the 32-dimensional second-order interaction features are processed to obtain 32-dimensional global adaptation intermediate features under the normalized adjacency matrix. The 32-dimensional global adaptation intermediate features are aggregated by the fourth layer GCN to obtain 32-dimensional full-scene high-order interaction features, including 32-dimensional full-scene high-order interaction features. In the third stage, the dynamic feature calibration coefficients of the third stage are loaded to calibrate the features of the third stage nodes to six centroids: lane keeping, left lane change, right lane change, deceleration, acceleration, and stopping. The calibrated target vehicle features are then input into the classification head to obtain the full-scene intent recognition results.

[0015] Based on the aforementioned technical means, the homogeneity judgment criteria are further relaxed, the complex connections related to multi-vehicle game in all scenarios are strengthened, and irrelevant noise interference is weakened, making the graph structure more consistent with the real topological distribution of global interactions within a 50-meter range; the numerical instability caused by differences in node degree is eliminated, ensuring the smoothness and convergence stability of feature values ​​during subsequent GCN aggregation; the learned car-following and lane-changing basic interaction capabilities are fully preserved, avoiding the forgetting problem caused by retraining deep networks; at the same time, the basic features are semantically aligned under the globally optimized graph structure, laying a solid foundation for the extraction of higher-order information in the fourth layer; the complex game information of multi-vehicle intertwining in all scenarios is covered, long-distance dependencies are encoded, and the feature expression ability of the model under complex conditions such as congestion and ramp merging is significantly improved; small step calibration further increases the feature spacing of the six types of intents while ensuring feature stability, improving intra-class compactness and inter-class discriminability; and in conjunction with the classification head, high-precision prediction of driving intent in all scenarios is achieved to meet the real-time decision-making needs of vehicles.

[0016] Furthermore, the processed target vehicle features are input into the intent classifier to obtain the driving intent output by the intent classifier, including: Extract the row vector corresponding to the target vehicle from the 32-dimensional full-scene high-order interaction features, and denote it as the processed target vehicle feature. The processed target vehicle features are input into the intent classifier to obtain the driving intent output by the intent classifier.

[0017] Based on the aforementioned technical means and high-level interactive features with high distinguishability, the driving intentions of the target vehicle in all scenarios (keeping the lane, changing lanes left / right, decelerating, accelerating, and stopping) are accurately output, meeting the high-precision requirements of real-time in-vehicle decision-making.

[0018] A vehicle driving intention prediction device, applied in a vehicle, the prediction device comprising: The first processing unit is used to select vehicles within the target distance as graph nodes with the target vehicle as the center, and extract multi-dimensional motion features of each vehicle to form a node attribute matrix. The first processing unit is used to traverse all vehicle pairs. If the distance between vehicle pairs is less than a preset distance, an edge is constructed, the edge weight corresponding to the edge is calculated, and an initial adjacency matrix is ​​generated. The second processing unit is used to perform multi-stage progressive graph self-enhancement processing based on the node attribute matrix and the initial adjacency matrix. In each stage, the corresponding number of layers of the graph convolutional network (GCN) and the adaptation parameters are loaded to obtain the processed target vehicle features. The prediction unit is used to input the processed target vehicle features into the intent classifier to obtain the driving intent output by the intent classifier.

[0019] A vehicle includes a processor and a memory, the memory storing a computer program executable on the processor, the processor executing the computer program to implement any of the methods described above.

[0020] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the method described in any of the preceding claims.

[0021] A computer program product includes a computer program or instructions that, when executed by a processor, implement the method described in any of the preceding claims.

[0022] The beneficial effects of this application are as follows: Vehicles within the target distance are selected as graph nodes centered on the target vehicle; multi-dimensional motion features are extracted from each vehicle to form a node attribute matrix; all vehicle pairs are traversed, and if the distance between vehicle pairs is less than a preset distance, an edge is constructed, and the edge weight is calculated to generate an initial adjacency matrix; based on the node attribute matrix and the initial adjacency matrix, a multi-stage progressive graph self-enhancement process is performed, loading a corresponding number of layers of a Graph Convolutional Network (GCN) and adaptation parameters at each stage to obtain the processed target vehicle features; the processed target vehicle features are input into an intent classifier to obtain the driving intent output by the intent classifier. This application can perform multi-stage progressive graph self-enhancement processing, progressively extracting interactive semantics from basic to complex scenarios, significantly improving the discriminative power and robustness of features, and enhancing prediction accuracy. Attached Figure Description

[0023] Figure 1 This application provides a schematic diagram of the implementation process of a vehicle driving intention prediction method. Figure 1 ; Figure 2 This application provides a schematic diagram of the implementation process of a vehicle driving intention prediction method. Figure 2 ; Figure 3This is a schematic diagram of a vehicle driving intention prediction process based on progressive graph self-reinforcement provided in the embodiments of this application; Figure 4 This is a schematic diagram of a progressive graph self-enhancement algorithm framework provided in the embodiments of this application; Figure 5 This is a schematic diagram of the overall process of predicting vehicle driving intention based on progressive graph self-reinforcement provided in the embodiments of this application; Figure 6 This is a schematic diagram of the structure of a vehicle driving intention prediction device provided in the embodiments of this application; Figure 7 This is a schematic diagram of the hardware entity of a vehicle provided in the embodiments of this application. Detailed Implementation

[0024] The embodiments of this application will be described below with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be understood that the preferred embodiments are only for illustrating this application and are not intended to limit the scope of protection of this application.

[0025] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. Therefore, the drawings only show the components related to this application and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0026] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0027] In the following description, the terms "first, second, third" are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0028] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0029] In the implementation of this application, the collection and processing of relevant data should strictly comply with the requirements of relevant laws and regulations, obtain the informed consent or separate consent of the personal information subject, and carry out subsequent data use and processing within the scope of laws and regulations and the authorization of the personal information subject.

[0030] The method provided in this application can be executed by an electronic device, which can be a vehicle, laptop, tablet, desktop computer, mobile device (e.g., mobile phone, personal digital assistant), or a server. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. The following description uses a vehicle as the executing entity.

[0031] The technical solutions in the embodiments of this application will now be clearly and completely described with reference to the accompanying drawings.

[0032] Figure 1 This is a schematic diagram illustrating the implementation process of a vehicle driving intention prediction method provided in an embodiment of this application, as shown below. Figure 1 As shown, the prediction method includes steps S11 to S14, wherein: Step S11: Select vehicles within the target distance as graph nodes, centered on the target vehicle, and extract multi-dimensional motion features for each vehicle to form a node attribute matrix.

[0033] In this embodiment, the target distance can be set based on the real-time driving scenario of the vehicle or based on historical driving scenarios. The target distance can be set to 50 meters, and the multi-dimensional range includes, but is not limited to, 10 dimensions; this application does not impose specific limitations on this.

[0034] As an example, taking the target vehicle as the center, all surrounding vehicles within a 50-meter radius are selected as graph nodes. The position, speed, acceleration, and other 10-dimensional motion features of each vehicle are extracted and concatenated into a graph with [dimensionality missing]. ×10 node attribute matrix ,in, Each line contains features such as the vehicle's position, speed, acceleration, and heading angle. It is the total number of nodes in the graph where vehicles within a 50-meter radius are used as nodes; This represents the total number of feature dimensions for a single vehicle node. In this way, the relevant interaction objects of the target vehicle are accurately identified, while interference from irrelevant vehicles at a certain distance is filtered out, providing clean, raw feature input for subsequent graph computation.

[0035] As an example, the 10-dimensional motion features include: position information: the vehicle's lateral and longitudinal relative coordinates (2D); motion information: the vehicle's absolute velocity, longitudinal velocity, lateral velocity, and acceleration (4D); heading information: the vehicle's heading angle and heading angular velocity (2D); and size information: the vehicle's length and width (2D). The extraction of multi-dimensional motion features for each vehicle can be set according to actual needs, and this application does not impose specific limitations on this comparison.

[0036] Step S12: Traverse all vehicle pairs. If the distance between vehicle pairs is less than the preset distance, construct an edge and calculate the edge weight corresponding to the edge to generate an initial adjacency matrix.

[0037] In this embodiment, the preset distance The settings can be based on real-time vehicle driving scenarios or historical vehicle driving scenarios. The preset distance includes, but is not limited to, Euclidean distance, such as 30 meters; this application does not specifically limit this.

[0038] As an example, iterate through all selected vehicle node pairs, and establish a connection edge if the distance between the two vehicles is less than 30 meters. Edge weights are calculated using the Gaussian kernel formula. Generate the initial adjacency matrix In this way, the physical interaction between vehicles can be intuitively expressed using topological relationships, providing a basic graph structure support for the subsequent aggregation of neighborhood information in the GCN.

[0039] In a feasible scenario of constructing a traffic scene graph, the vehicle driving scene is constructed as an original attribute graph. , where: node set Each node represents a vehicle (including the target vehicle and surrounding vehicles). Adjacency matrix formed by edge sets Based on the relationship between spatial distance and lane, when the vehicle With vehicles Spatial distance less than threshold At that time, construct the edge The edge weight is In other words, when the spatial distance between the two vehicles is less than a set threshold... At that time, Effective, used to characterize the degree of correlation between two vehicles in a traffic scenario.

[0040] Node attribute matrix Each line contains features such as the vehicle's position, speed, acceleration, and heading angle.

[0041] Step S13: Based on the node attribute matrix and the initial adjacency matrix, perform multi-stage progressive graph self-enhancement processing. In each stage, load the corresponding number of layers of the graph convolutional network (GCN) and the adaptation parameters to obtain the processed target vehicle features.

[0042] In this embodiment, a multi-stage progressive graph self-enhancement processing method is adopted to progressively extract the full-scene interaction semantics from basic following to complex game theory. The multi-stage process includes, but is not limited to, three stages.

[0043] As an example, the GCN and dedicated adaptation parameters for the corresponding layers are loaded in three progressive stages. First-order neighbor information is aggregated from the second-layer GCN, second-order neighbor information is expanded from the third-layer GCN, and finally, the fourth-layer GCN completes the aggregation of high-order interactions across the entire graph. The graph structure and feature distribution are optimized step by step to output the final high-order features of the target vehicle. In this way, the problem of oversmoothing that is prone to occur in single-stage deep processing (such as single-stage deep GCN or GNN processing) is avoided. The interaction semantics of the whole scene, from basic car following to complex game, are extracted layer by layer, which greatly improves the discriminativeness and robustness of the features.

[0044] Step S14: Input the processed target vehicle features into the intent classifier to obtain the driving intent output by the intent classifier.

[0045] In this embodiment, the intent classifier comprises a fully connected layer and a Softmax layer. This classifier maps the processed target vehicle features to a probability distribution of six types of driving intents, and takes the maximum value as the final prediction result. Thus, based on highly discriminative interaction features, the driving intent of the target vehicle (such as changing lanes to the left or decelerating) is accurately output, meeting the prediction accuracy requirements of real-time in-vehicle scenarios.

[0046] This application provides a method for predicting vehicle driving intentions, applied to vehicles. The method includes: selecting vehicles within a target distance as graph nodes, and extracting multi-dimensional motion features from each vehicle to form a node attribute matrix; traversing all vehicle pairs, constructing edges if the distance between vehicle pairs is less than a preset distance, and calculating the edge weights to generate an initial adjacency matrix; based on the node attribute matrix and the initial adjacency matrix, performing multi-stage progressive graph self-enhancement processing, loading a corresponding number of layers of a Graph Convolutional Network (GCN) and adaptation parameters at each stage to obtain processed target vehicle features; and inputting the processed target vehicle features into an intention classifier to obtain the driving intention output by the intention classifier. This application can perform multi-stage progressive graph self-enhancement processing, progressively extracting interactive semantics from basic to complex scenarios, significantly improving the discriminative power and robustness of features, and enhancing prediction accuracy.

[0047] In some embodiments, step S13 performs multi-stage progressive graph self-enhancement processing based on the node attribute matrix and the initial adjacency matrix. At each stage, a graph convolutional network (GCN) of the corresponding number of layers and adaptation parameters are loaded to obtain the processed target vehicle features, including, for example... Figure 2 Steps S131 to S133 shown are as follows: Step S131: Based on the node attribute matrix and the initial adjacency matrix, load the 2-layer GCN and the adaptation parameters of the first stage in the first stage for processing, and output the first-order interaction features and car-following recognition results.

[0048] In this embodiment, the adaptation parameters for the first stage include a homogeneity enhancement threshold, a homogeneity suppression threshold, an enhancement coefficient, a suppression coefficient, and a dynamic feature calibration coefficient. As an example, the homogeneity enhancement threshold for the first stage... =0.7, Homogeneity Inhibition Threshold =0.3, enhancement coefficient =0.3, inhibition coefficient =0.2, Dynamic Feature Calibration Coefficient =0.15.

[0049] As an example, loading a 2-layer GCN with the first-stage homogeneity enhancement threshold... =0.7, Homogeneity Inhibition Threshold =0.3, enhancement coefficient =0.3, inhibition coefficient =0.2, Dynamic Feature Calibration Coefficient =0.15; Based on the original node attribute matrix and initial adjacency matrix, only the first-order direct neighbor information of the target vehicle is aggregated to complete large-step feature calibration and graph structure optimization. In this way, basic car-following interaction features are quickly extracted, and 32-dimensional first-order interaction features and binary classification car-following recognition results are output, avoiding the premature oversmoothing problem of deep networks.

[0050] Step S132: Based on the first-order interaction features and the updated adjacency matrix obtained in the first stage, the 3-layer GCN and the adaptation parameters of the second stage are loaded in the second stage for processing, and the second-order interaction features and lane change recognition results are output.

[0051] In this embodiment, the adaptation parameters for the second stage include a homogeneity enhancement threshold, a homogeneity suppression threshold, an enhancement coefficient, a suppression coefficient, and a dynamic feature calibration coefficient. As an example, the homogeneity enhancement threshold for the second stage... =0.6, Homogeneity Inhibition Threshold =0.4, enhancement coefficient =0.4, inhibition coefficient =0.25, Dynamic Feature Calibration Coefficient =0.1.

[0052] As an example, the second stage reuses the weights of the first two GCN layers, adds a third GCN layer, and loads the homogeneity enhancement threshold of the second stage. =0.6, Homogeneity Inhibition Threshold =0.4, enhancement coefficient =0.4, inhibition coefficient =0.25, Dynamic Feature Calibration Coefficient =0.1; Based on the first-order interaction features and the adjacency matrix updated in the first stage, the interaction range is extended to second-order neighboring vehicles, and the mid-step feature calibration is completed. In this way, while preserving the car-following semantics, the motion information of the vehicles in front and behind in the adjacent lane is fused to output 32-dimensional second-order interaction features and three-class lane change recognition results, balancing convergence speed and feature stability.

[0053] Step S133: Based on the second-order interaction features and the second-stage updated adjacency matrix, the third stage loads the 4-layer GCN and the third-stage adaptation parameters for processing, and outputs the high-order interaction features and the full-scene intent recognition results; wherein, the processed target vehicle features include high-order interaction features.

[0054] In this embodiment, the adaptation parameters for the third stage include a homogeneity enhancement threshold, a homogeneity suppression threshold, an enhancement coefficient, a suppression coefficient, and a dynamic feature calibration coefficient. As an example, the homogeneity enhancement threshold for the third stage... =0.5, Homogeneity Inhibition Threshold =0.5, enhancement coefficient =0.5, inhibition coefficient =0.3, Dynamic Feature Calibration Coefficient =0.05.

[0055] As an example, the third stage reuses the weights of the first three GCN layers, adds a fourth GCN layer, and loads the homogeneity enhancement threshold of the third stage. =0.5, Homogeneity Inhibition Threshold =0.5, enhancement coefficient =0.5, inhibition coefficient =0.3, Dynamic Feature Calibration Coefficient =0.05; Based on second-order interaction features and the second-stage second-updated adjacency matrix, high-order interaction aggregation of all vehicles within 50 meters of the entire map is opened, and fine-grained feature calibration with small steps is completed. In this way, multi-vehicle game information covering the entire scene is covered, outputting 32-dimensional high-order interaction features and 6-class full-scene intent recognition results. Finally, the row vector corresponding to the target vehicle is extracted as the processed target vehicle feature, realizing high-precision intent prediction in complex scenes.

[0056] In some embodiments, step S131, based on the node attribute matrix and the initial adjacency matrix, loads the two-layer GCN and the adaptation parameters of the first stage for processing, and outputs the first-order interaction features and car-following recognition results, can be achieved through the following steps: A11. Based on the node attribute matrix, two layers of GCN are loaded in the first stage. The first layer of GCN maps the multidimensional motion features to 64-dimensional motion features, and the second layer of GCN maps the 64-dimensional motion features to 32-dimensional motion features. Among them, the first-order interaction features are 32-dimensional motion features.

[0057] In this embodiment, the 10-dimensional original node attribute matrix is ​​input into a two-layer GCN. The first-layer GCN aggregates first-order neighbor vehicle information, mapping the 10-dimensional features to 64 dimensions. The second-layer GCN further aggregates the 64-dimensional features of the first-order neighbor vehicles, compressing them to obtain 32-dimensional first-order interaction features. In this way, only the car-following interaction information of directly adjacent vehicles is focused on, avoiding the introduction of irrelevant noise and quickly completing the extraction of basic interaction semantics.

[0058] A12. In the first stage, load the homogeneity enhancement threshold, homogeneity suppression threshold, enhancement coefficient, and suppression coefficient of the first stage, and dynamically update the initial adjacency matrix to obtain the updated adjacency matrix.

[0059] In this embodiment, first-stage specific parameters, such as the first-stage homogeneity enhancement threshold, are loaded. =0.7, Homogeneity Inhibition Threshold =0.3, enhancement coefficient =0.3, inhibition coefficient =0.2, calculate the homogeneity index of all vehicle pairs, enhance the weight of high similarity edges and suppress the weight of low similarity edges to generate an optimized adjacency matrix. In this way, the connection of car-following related edges is strengthened and the irrelevant interference edges are weakened, making the graph structure more consistent with the real interaction logic of car-following scenarios.

[0060] A13. In the first stage, load the dynamic feature calibration coefficient of the first stage and calibrate the node features of the first stage to the centroids of the following and non-following categories to obtain the following recognition result.

[0061] In this embodiment, the first-stage dynamic feature calibration coefficients are loaded. =0.15, rapidly pulling the 32-dimensional node features closer to the centroids of "carrying" and "non-carrying" classes, and then outputting the final carrying recognition result through the binary classification head. In this way, the large stride allows the dispersed features to quickly converge to the two class centers, significantly improving the convergence speed and recognition accuracy of carrying binary classification.

[0062] In this embodiment, the first-order interaction feature is the 32-dimensional motion feature finally output by the two-layer GCN in the first stage. It is a fixed-dimensional latent vector that retains only single-hop vehicle interaction information after first-order neighbor aggregation. The first-stage node feature is a collective term for the node features of all intermediate states in the entire first-stage process. It includes the intermediate features output layer by layer by the two-layer GCN, the final 32-dimensional first-order interaction feature, and all node features before and after the first-stage dynamic feature calibration.

[0063] In some embodiments, step S132, based on the first-order interaction features and the updated adjacency matrix obtained in the first stage, loads the 3-layer GCN and the adaptation parameters of the second stage in the second stage for processing, and outputs the second-order interaction features and lane-change recognition results. This can be achieved through the following steps: B11. In the second stage, load a 3-layer GCN. By reusing the weights of the first and second layers of GCN, process the 32-dimensional motion features to obtain the reconstructed features under the updated adjacency matrix.

[0064] In this embodiment, the weights of the first two GCN layers that have been trained and converged in the first stage are loaded, and the 32-dimensional first-order interaction features output by the first stage are used as input to complete feature reconstruction with the support of the adjacency matrix updated in the first stage. In this way, the learned basic interaction capabilities are fully preserved, avoiding the forgetting problem caused by training a 3-layer GCN from scratch, and quickly adapting to the graph structure of lane-changing scenarios.

[0065] B12. The reconstructed features are aggregated through the third-layer GCN to obtain 32-dimensional second-order interaction features, where the second-order interaction features are 32-dimensional second-order interaction features.

[0066] In this embodiment, the newly added third-layer GCN performs neighborhood aggregation on the 32-dimensional reconstructed features, expands the interactive receptive field to second-order neighboring vehicles, and finally outputs 32-dimensional second-order interactive features. In this way, the motion information of the vehicles in front and behind in the adjacent lane is fused to encode the multi-vehicle game relationship in the lane-changing scenario, providing high-order semantic support for subsequent lane-changing recognition.

[0067] B13. In the second stage, load the homogeneity enhancement threshold, homogeneity suppression threshold, enhancement coefficient, and suppression coefficient of the second stage, and dynamically update the updated adjacency matrix to obtain the second-updated adjacency matrix.

[0068] In this embodiment, second-stage specific parameters, such as the second-stage homogeneity enhancement threshold, are loaded. =0.6, Homogeneity Inhibition Threshold =0.4, enhancement coefficient =0.4, inhibition coefficient =0.25, calculate the homogeneity index, adjust the edge weights of the adjacency matrix after the first stage update, and generate the adjacency matrix after the second update; in this way, the lateral vehicle connection related to lane changing is strengthened, the graph topology is further optimized, and the interaction distribution of lane changing scenarios is adapted.

[0069] B14. In the second stage, the dynamic feature calibration coefficients of the second stage are loaded to calibrate the centroids of the second stage node features to maintain the lane, left lane change, and right lane change, so as to obtain the lane change recognition results.

[0070] In this embodiment, the second-stage dynamic feature calibration coefficients are loaded. =0.1, calibrate the node features to the centroids of three categories: lane keeping, left lane change, and right lane change, and then output the lane change recognition result through the three-class classification head; in this way, the convergence speed and feature stability are balanced, feature oscillation caused by large step size is avoided, and the recognition accuracy of lane change three-class classification is greatly improved.

[0071] In this embodiment, the second-order interaction feature is the 32-dimensional feature finally output by the third-layer GCN added in the second stage, which completes the aggregation of vehicle interactions within a two-hop range and integrates indirect interaction information from adjacent lanes and vehicles in front and behind. The second-stage node feature is the set of node features for the entire second-stage process, including the reconstructed features obtained by reusing the weights of the first two layers of GCN, the original features output by the third-layer GCN, and all node features before and after the second-stage dynamic feature calibration.

[0072] In some embodiments, step S133, based on the second-order interaction features and the second-updated adjacency matrix obtained in the second stage, loads the 4-layer GCN and the adaptation parameters of the third stage in the third stage for processing, outputting high-order interaction features and full-scene intent recognition results; wherein, the processed target vehicle features include high-order interaction features, which can be achieved through the following steps: C11. In the third stage, load the homogeneity enhancement threshold, homogeneity suppression threshold, enhancement coefficient, and suppression coefficient of the third stage, and dynamically update the adjacency matrix after the second update to obtain the adjacency matrix after the third update.

[0073] In this embodiment, third-stage specific parameters, such as the third-stage homogeneity enhancement threshold, are loaded. =0.5, Homogeneity Inhibition Threshold =0.5, enhancement coefficient =0.5, inhibition coefficient =0.3, based on the second-stage output of the adjacency matrix after the second-stage update. Calculate the homogeneity index of the current node's features. Boost the weights of high-similarity edges and suppress the weights of low-similarity edges, generating the original adjacency matrix after three updates. Thus, by further relaxing the homogeneity criteria, strengthening the complex connections related to multi-vehicle game in all scenarios, and weakening irrelevant noise interference, the graph structure can better fit the real topological distribution of global interactions within a 50-meter range.

[0074] C12. Normalize the adjacency matrix after three updates to obtain the normalized adjacency matrix.

[0075] In this embodiment, the dynamically updated Add self-loops, calculate the matrix, and perform symmetric normalization. This yields the normalized adjacency matrix used for GCN convolution operations. This eliminates the numerical instability caused by differences in node degree, ensuring the smoothness and convergence stability of feature values ​​during subsequent GCN aggregation.

[0076] C13. In the third stage, load a 4-layer GCN. By reusing the weights of the first, second, and third layers of GCN, process the 32-dimensional second-order interaction features to obtain 32-dimensional globally adapted intermediate features under the normalized adjacency matrix.

[0077] In this embodiment, the weights of the first three converged layers of the GCN, trained in the first and second stages, are loaded, along with the 32-dimensional second-order interaction features output from the second stage. As input, in the normalized adjacency matrix Supported by this, the features are reconvolved and reconstructed, outputting 32-dimensional globally adapted intermediate features. In this way, the learned basic interaction capabilities of car following and lane changing are fully preserved, avoiding the forgetting problem caused by retraining deep networks; at the same time, the basic features are semantically aligned under the globally optimized graph structure, laying a solid foundation for the extraction of higher-order information in the fourth layer.

[0078] C14. The 32-dimensional global adaptation intermediate features are aggregated through the fourth layer GCN to obtain 32-dimensional full-scene high-order interaction features, including 32-dimensional full-scene high-order interaction features.

[0079] In this embodiment, the newly added fourth-layer GCN performs full-map neighborhood aggregation on the 32-dimensional global adaptation intermediate features, expanding the interaction receptive field to all vehicles within 50 meters, and finally outputting 32-dimensional full-scene high-order interaction features. In this way, the complex game information involving multiple vehicles in all scenarios is covered, and long-distance dependencies are encoded, significantly improving the model's feature representation ability under complex conditions such as congestion and ramp merging.

[0080] C15. In the third stage, the dynamic feature calibration coefficients of the third stage are loaded to calibrate the node features of the third stage to six centroids: lane keeping, left lane changing, right lane changing, deceleration, acceleration, and stopping. The calibrated target vehicle features are then input into the classification head to obtain the full-scene intent recognition results.

[0081] In this embodiment, the dynamic feature calibration coefficient in loading stage 3 is... =0.05, fine-tuning and calibrating the node features to six centroids: lane keeping, left / right lane changing, deceleration, acceleration, and stopping; then, from the calibrated... The target vehicle features are extracted and input into a fully connected classification head (Linear+Softmax), which outputs the probability distribution of six types of driving intentions and the final result. Thus, small-step calibration, while ensuring feature stability, further increases the feature spacing between the six types of intentions, improving intra-class compactness and inter-class discriminability; combined with the classification head, it achieves high-precision prediction of driving intentions across all scenarios, meeting the real-time decision-making needs of in-vehicle systems.

[0082] In this embodiment, the high-order interaction features are the 32-dimensional full-scene features finally output by the fourth-layer GCN added in the third stage. These features aggregate high-order vehicle interactions with three or more hops, covering global interaction information for multi-vehicle games and complex traffic flows. The third-stage node features are the set of node features throughout the entire third-stage process. They include the 32-dimensional globally adapted intermediate features obtained by reusing the weights of the first three layers of GCN, the original features output by the fourth-layer GCN, and all node features before and after the third-stage dynamic feature calibration.

[0083] In some embodiments, step S14 inputs the processed target vehicle features into the intent classifier to obtain the driving intent output by the intent classifier. This can be achieved through the following steps: Extract the row vector corresponding to the target vehicle from the 32-dimensional full-scene high-order interaction features, and denote it as the processed target vehicle feature. The processed target vehicle features are input into the intent classifier to obtain the driving intent output by the intent classifier.

[0084] In this embodiment, the 32-dimensional vector corresponding to the target vehicle is indexed from the N×32-dimensional full-scene high-order interaction feature matrix output from the third stage and used as a separate feature specific to the target vehicle. This filters out redundant features from other irrelevant vehicles, accurately extracting the full-scene interaction semantics belonging only to the target vehicle, providing clean and dedicated input for subsequent classification. Furthermore, the 32-dimensional target vehicle feature is fed into a 6-class intent classifier composed of a fully connected layer and Softmax activation, mapping it to a probability distribution of 6 types of driving intents. The maximum probability is taken as the final output driving intent. This achieves the following effect: based on highly discriminative high-order interaction features, the full-scene driving intents of the target vehicle (lane keeping, left / right lane change, deceleration, acceleration, and stopping) are accurately output, meeting the high-precision requirements of real-time in-vehicle decision-making.

[0085] Figure 3 This is a schematic diagram of a vehicle driving intention prediction process based on progressive graph self-reinforcement provided in the embodiments of this application, such as... Figure 3 As shown, the prediction method includes steps S21 to S24, wherein: Step S21: Construct a traffic scenario map.

[0086] Here, the vehicle driving scenario is constructed as an original attribute graph. ,in: Node set Each node represents a vehicle (including the target vehicle and surrounding vehicles). Adjacency matrix formed by edge sets Based on the relationship between spatial distance and lane, when the vehicle With vehicles Spatial distance less than threshold At that time, construct the edge The edge weight is In other words, when the spatial distance between the two vehicles is less than a set threshold... At that time, Effective, used to characterize the degree of correlation between two vehicles in a traffic scenario.

[0087] Node attribute matrix Each line contains features such as the vehicle's position, speed, acceleration, and heading angle.

[0088] Step S22: Construct a graph encoder.

[0089] Here, a graph encoder consisting of multiple graph convolutional layers is constructed to extract feature representations of vehicle nodes: ; ; ; in, That is, the initial input Original node attribute matrix , This refers to the number of GCN layers. For the first The node feature matrix output by the layer graph convolutional layer. For the first The node feature matrix output by the layer graph convolutional layer. For the first The trainable weight matrix of a layer GCN The normalized adjacency matrix, It is a diagonal matrix. To add a self-connected adjacency matrix, It is a non-linear activation function. It is an identity matrix.

[0090] Step S23: Construct a progressive graph self-reinforcing algorithm framework.

[0091] Figure 4 This is a schematic diagram of a progressive graph self-enhancement algorithm framework provided in the embodiments of this application, such as... Figure 4 As shown, a progressive enhancement framework is constructed, consisting of four parts: a learning scheduler, a graph structure self-enhancement module, a node feature optimization module, and a dynamic balance controller.

[0092] (1) Learning scheduler: The training process is divided into three progressive phases, each using a different GCN depth and training objective, as shown in Table 1:

[0093] Table 1 As an example, in Table 1 It is 2. It is 3. The value is 4.

[0094] In practical applications, the learning scheduler is dynamically controlled by the feature convergence and class separation of the nodes. Let the node feature matrix of the current stage be... , This represents the total number of vehicle nodes in the current graph, corresponding to the total number of individual vehicles participating in the interaction within the graph. This represents the dimension of the feature vector output by each vehicle node after encoding by the current GCN layer, which is the length of the final feature representation obtained by a single node; here, the class separation metric is: ; in, This represents the total number of vehicle intent categories in a driving scenario (e.g., 6, where intent categories include lane keeping, left lane change, right lane change, deceleration, acceleration, and stopping). and For two different vehicle intent categories, and For the feature centroids of different intention categories. When The improvement is less than the threshold for three consecutive epochs. At that time, the trigger phase progresses.

[0095] (2) Graph structure self-enhancing module: The connection strength of the graph structure is dynamically adjusted to strengthen intra-class connections and suppress inter-class noise propagation.

[0096] Homogeneity assessment: Based on the node characteristics at the current stage, calculate the node... With nodes Homogeneity index: ; in, A higher value indicates that the two cars have more similar driving patterns. is the feature space scale hyperparameter. Represents a node eigenvectors, Represents a node eigenvectors, and Used to calculate feature similarity between nodes. This represents the feature similarity, or homogeneity index, between two nodes. Understandably, here... It is the first The original hidden layer feature vectors of each vehicle node after extraction by GCN, but before dynamic feature calibration, are... It is the first The original hidden layer feature vectors of each vehicle node after extraction by GCN, but before dynamic feature calibration. Understandably, the first-stage node features, second-stage node features, and third-stage node features belong to... In terms of their concrete forms at different training stages, the three correspond to the hidden vectors of the current layer nodes output by GCNs of different depths in the first, second, and third stages, respectively.

[0097] Here, the connection strength is dynamically adjusted: the adjacency matrix is ​​dynamically updated based on the homogeneity index. ; ; in, To enhance the coefficient, The suppression coefficient, To enhance the threshold for homogeneity, This is the homogeneity inhibition threshold.

[0098] (3) Node feature optimization module: It guides node features to converge toward the class centroid, while simultaneously forcibly separating different classes in the feature space.

[0099] Dynamic feature calibration: This process moves node features closer to the centroid of their respective intent categories. ; ; in, Belongs to a specific intention category The set of nodes, For set The traversal index condition in the text refers to traversing all elements belonging to a specific intent category. The vehicle nodes are used to extract their feature vectors, and then... Calculate the average of their eigenvectors to determine the centroid of the specific intention. ,here, For category The centroid vector. Then, using the formula Original features (The original feature to be calibrated) is pulled closer to the centroid to obtain the calibration feature. The above. This refers to the calibration coefficient (also known as the dynamic characteristic calibration coefficient).

[0100] Here, the class centroid separation constraint forces the feature centroids of different intent categories to maintain a distance. ; in, The minimum separation interval hyperparameter, and For the feature centroids of different intent categories, Refers to the loss due to the separation constraint of the class center.

[0101] (4) Dynamic balance controller: The model's status is monitored in real time by a dynamic balancing controller, which dynamically adjusts the propulsion progress and enhancement intensity. Oversmoothing monitoring metrics are defined. The ratio of intra-class compactness to inter-class dispersion: ; in, ; in, ; in, The trace of the matrix, The scatter matrix is ​​the intra-class scatter matrix. The inter-class scatter matrix, Here, the centroids of the global class features are... =6 (including lane keeping, left lane change, right lane change, deceleration, acceleration, and stopping). For category indexing, iterate from category 1 to category 2. All intent categories of the class. Belongs to the A set of vehicle nodes representing a class intent. Represents a node The eigenvector of , i.e., the eigenvector of . The current feature vector of each vehicle node is the hidden layer feature representation extracted by GCN. It is the first The feature centroid vector of the class intent. For the first The number of nodes contained in the class intent.

[0102] Here, when Higher than the preset threshold If the model is determined to be in an oversmoothing risk state, the following adjustment actions will be automatically executed: The number of GCN layers is reduced by one to limit excessive information mixing. This adjustment does not change the model training phase; Increase the weight coefficient of the class center separation constraint ; The trigger graph structure self-enhancement module corrects the node connection weights.

[0103] In this embodiment, the dynamic balance controller calculates... To evaluate the quality of the feature space in real time. If Higher than the preset threshold This indicates that the class is not compact enough or that the classes are not sufficiently separated (resulting in a risk of excessive smoothing). This will automatically trigger corresponding adjustments, such as automatically reducing the number of GCN layers (i.e., reducing the number of neighbor information aggregations) and increasing... (i.e., enhance inter-class separation loss) and correct graph structure to restore the discriminative power of features.

[0104] Step S24: Intent prediction and model training.

[0105] The target vehicle features after progressive graph self-enhancement processing Input to a fully connected layer and In the intent classifier constructed from functions, the driving intent category is predicted: ; The loss function for predicting the target vehicle's driving intention category is cross-entropy loss: ; in, For the number of intent categories, For the target vehicle node Authentic labels coding.

[0106] Total loss function of model training The calculation is as follows: ; ; in, For feature calibration loss, Weighting coefficients for feature calibration loss These are the weighting coefficients for the class center separation constraint. This represents the total number of all vehicle nodes participating in the calculation within the current training batch.

[0107] This application provides a vehicle driving intention prediction method based on progressive graph self-reinforcement. By constructing a progressive training framework from easy to difficult, dynamically adjusting the learning strategy of the graph structure and the node feature optimization process, it solves the technical problems of oversmoothing and training difficulties that traditional methods easily encounter when processing high-order interactive information, thereby improving the accuracy and robustness of driving intention prediction.

[0108] Figure 5 This is a schematic diagram of the overall process of predicting vehicle driving intention based on progressive graph self-reinforcement provided in the embodiments of this application, such as... Figure 5 As shown, the prediction method includes steps S31 to S35, wherein: Step S31: Data acquisition and preprocessing.

[0109] As an example, the vehicle's status data and that of surrounding vehicles are collected using an in-vehicle forward-facing camera, millimeter-wave radar, and Global Positioning System (GPS). The collected data includes: Vehicle status: position coordinates, speed, acceleration, heading angle; Surrounding vehicle status: relative position, relative speed, lane position (same lane / left lane / right lane); All data were preprocessed by normalization and time alignment, with a sampling frequency of 10Hz, and 3 seconds (30 frames) of historical data were used as input.

[0110] Step S32: Traffic scene graph construction.

[0111] As an example, taking the target vehicle as the center, vehicles within a 50-meter radius are selected as graph nodes, including the target vehicle, the vehicle in front, the vehicle behind, the vehicle to the left front, the vehicle to the left rear, the vehicle to the right front, the vehicle to the right rear, and vehicles further ahead. Node features include: normalized longitudinal position, lateral position, longitudinal velocity, lateral velocity, longitudinal acceleration, lateral acceleration, heading angle, rate of change of heading angle, vehicle type code, and lane position code. Feature dimensions: .

[0112] Here, the edge construction rule is as follows: when the Euclidean distance between two vehicles is less than 30 meters, an edge is constructed between the two vehicles. The edge weight calculation formula is as follows: ; in, For vehicles With vehicles The Euclidean distance between them Take 15 meters.

[0113] Step S33: Construct a graph encoder.

[0114] As an example, the graph encoder uses a multi-layer graph convolutional network with the following parameter settings: Input dimensions: ; First layer Output dimensions: Used in stages 1 / 2 / 3; Second floor Output dimensions: Used in stages 1 / 2 / 3; Third layer Output dimensions: Used in phase 2 / 3; Fourth floor Output dimensions: Used in Phase 3; Non-linear activation function: .

[0115] Step S34: Construct a progressive graph self-reinforcing framework.

[0116] Here, we'll learn about scheduler settings: The overall parameter settings for the three stages are shown in Table 2:

[0117] Table 2 Here, the stage advancement threshold is set. ,if If the improvement is less than this value for three consecutive epochs, a phase advancement is triggered.

[0118] In a feasible scenario, an example of a graph structure self-reinforcing implementation is as follows: Set feature space scale This value primarily controls the rate of decay of the homogeneity index.

[0119] Taking Phase 2 as an example, when the homogeneity index of two vehicles... At the same time, increase its connection weight: When the homogeneity index of two vehicles At that time, suppress its connection weights: .

[0120] In a feasible scenario, an example of node feature optimization is implemented as follows: Dynamic feature calibration coefficient This represents the step size for the node feature to move towards the centroid, which is set to a large step size in stage 1. This accelerates the convergence of features to various centroids; in stage 2, a medium step size is set. To balance convergence speed and stability; a small step size is set in stage 3. Make fine adjustments to avoid fluctuations.

[0121] Taking the target vehicle belonging to the "left lane change" category in Phase 2 as an example, its category centroid... It is obtained by averaging the features of all samples in this category. The calibration operation brings the target vehicle features closer to the centroid of this category: ; Set the minimum separation interval hyperparameter The class center separation constraint is calculated as follows: ; In a feasible scenario, an example of implementing a dynamic balance controller is as follows: Monitor the ratio of intra-class compactness to inter-class dispersion. Set oversmoothing risk threshold ,when If the value exceeds this threshold, it is considered an oversmoothing risk, and the system automatically performs the following: temporarily disables the computation of the highest-level GCN in the current stage; for example, in stage 2, only the first two layers of GCN are used for forward propagation. The weight coefficients of the class center separation constraint are adjusted. Increased by 50%. The trigger graph structure self-enhancement module corrects the node connection weights.

[0122] Step S35: Intent prediction.

[0123] Here, the target vehicle features will be processed using progressive graph self-enhancement. (Right now The corresponding target vehicle node The row vector, with dimensions 32, is input into the intent classifier to obtain the driving intent classification probability vector of the target vehicle. The intent classifier consists of a fully connected layer and The activation function consists of a fully connected layer with an input dimension of 32 and an output dimension of 6 (corresponding to 6 types of intents: left lane change, right lane change, straight ahead, deceleration, acceleration, and stopping).

[0124] if If so, the predicted driving intention of the target vehicle is to change lanes to the left.

[0125] Set the weight coefficients for the feature calibration loss. Weight coefficients of class center separation constraints Then the total loss function of the model is: The model uses the Adam optimizer, and the learning rate can be set to 0.001.

[0126] Figure 6 This is a schematic diagram of the structure of a vehicle driving intention prediction device provided in the embodiments of this application, such as... Figure 6 As shown, the hardware entity of the vehicle driving intention prediction device 600 includes: The first processing unit 601 is used to filter vehicles within the target distance as graph nodes with the target vehicle as the center, and extract multi-dimensional motion features of each vehicle to form a node attribute matrix. The first processing unit 601 is used to traverse all vehicle pairs. If the distance between vehicle pairs is less than a preset distance, an edge is constructed, the edge weight corresponding to the edge is calculated, and an initial adjacency matrix is ​​generated. The second processing unit 602 is used to perform multi-stage progressive graph self-enhancement processing based on the node attribute matrix and the initial adjacency matrix. In each stage, the corresponding number of layers of the graph convolutional network (GCN) and the adaptation parameters are loaded to obtain the processed target vehicle features. The prediction unit 603 is used to input the processed target vehicle features into the intent classifier to obtain the driving intent output by the intent classifier.

[0127] In some embodiments, the second processing unit 602 is configured to, based on the node attribute matrix and the initial adjacency matrix, load two layers of GCN and the adaptation parameters of the first stage for processing in the first stage, and output first-order interaction features and car-following recognition results; based on the first-order interaction features and the updated adjacency matrix obtained in the first stage, load three layers of GCN and the adaptation parameters of the second stage for processing in the second stage, and output second-order interaction features and lane-change recognition results; based on the second-order interaction features and the second-updated adjacency matrix obtained in the second stage, load four layers of GCN and the adaptation parameters of the third stage for processing in the third stage, and output high-order interaction features and full-scene intent recognition results; wherein the processed target vehicle features include high-order interaction features.

[0128] In some embodiments, the second processing unit 602 is used to load two layers of GCN in the first stage based on the node attribute matrix. The first layer of GCN maps the multidimensional motion features to 64-dimensional motion features, and the second layer of GCN maps the 64-dimensional motion features to 32-dimensional motion features, wherein the first-order interaction features are 32-dimensional motion features. In the first stage, the homogeneity enhancement threshold, homogeneity suppression threshold, enhancement coefficient, and suppression coefficient of the first stage are loaded to dynamically update the initial adjacency matrix to obtain the updated adjacency matrix. In the first stage, the dynamic feature calibration coefficient of the first stage is loaded to calibrate the node features of the first stage to the centroids of the car-following and non-car-following categories to obtain the car-following recognition result.

[0129] In some embodiments, the second processing unit 602 is used to load a 3-layer GCN in the second stage, process the 32-dimensional motion features by reusing the weights of the first and second layer GCNs to obtain reconstructed features under the updated adjacency matrix; aggregate the reconstructed features through the third layer GCN to obtain 32-dimensional second-order interaction features, wherein the second-order interaction features are 32-dimensional second-order interaction features; load the homogeneity enhancement threshold, homogeneity suppression threshold, enhancement coefficient and suppression coefficient of the second stage in the second stage to dynamically update the updated adjacency matrix to obtain the second-updated adjacency matrix; load the dynamic feature calibration coefficient of the second stage in the second stage to calibrate the second-stage node features to three centroids: lane keeping, left lane changing and right lane changing, to obtain lane changing recognition results.

[0130] In some embodiments, the second processing unit 602 is used to load the homogeneity enhancement threshold, homogeneity suppression threshold, enhancement coefficient, and suppression coefficient of the third stage in the third stage, dynamically update the adjacency matrix after the second update, and obtain the adjacency matrix after the third update; normalize the adjacency matrix after the third update to obtain the normalized adjacency matrix; load a 4-layer GCN in the third stage, and process the 32-dimensional second-order interaction features by reusing the weights of the first, second, and third layers of GCN to obtain 32-dimensional global adaptation intermediate features under the normalized adjacency matrix; aggregate the 32-dimensional global adaptation intermediate features through the fourth layer GCN to obtain 32-dimensional full-scene high-order interaction features, wherein the high-order interaction features include 32-dimensional full-scene high-order interaction features; load the dynamic feature calibration coefficient of the third stage in the third stage to calibrate the third stage node features to six centroids: lane keeping, left lane changing, right lane changing, deceleration, acceleration, and stopping, and input the calibrated target vehicle features into the classification head to obtain the full-scene intent recognition result.

[0131] In some embodiments, the prediction unit 603 is used to extract the row vector corresponding to the target vehicle from the 32-dimensional full-scene high-order interaction features, denoted as the processed target vehicle features; the processed target vehicle features are input into the intent classifier to obtain the driving intent output by the intent classifier.

[0132] The descriptions of the above device embodiments are similar to those of the above method embodiments, and have similar beneficial effects. For technical details not disclosed in the device embodiments of this application, please refer to the descriptions of the method embodiments of this application for understanding.

[0133] It should be noted that, in the embodiments of this application, if the above methods are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to related technologies, can be embodied in the form of software products. These software products are stored in a storage medium and include several instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), magnetic disks, or optical disks. Thus, the embodiments of this application are not limited to any specific hardware and software combination.

[0134] This application also provides a vehicle including a memory and a processor, the memory storing a computer program that can run on the processor, and the processor executing the computer program to implement any of the methods described above.

[0135] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method. The computer-readable storage medium can be transient or non-transient.

[0136] This application also provides a computer program product, which includes a computer program or instructions that, when executed by a processor, implement some or all of the steps in any of the above-described methods. The computer program product can be implemented specifically through hardware, software, or a combination thereof. In one optional embodiment, the computer program product is specifically embodied in a computer storage medium; in another optional embodiment, the computer program product is specifically embodied in a software product, such as a software development kit (SDK), etc.

[0137] It should be noted that, Figure 7 This is a schematic diagram of the hardware entity of a vehicle provided in the embodiments of this application, such as... Figure 7 As shown, the hardware components of the vehicle 700 include a processor 701, a communication interface 702, and a memory 703. The processor 701 typically controls the overall operation of the vehicle 700. The communication interface 702 enables the vehicle to communicate with other terminals or servers via a network.

[0138] The memory 703 is configured to store instructions and applications executable by the processor 701, and can also cache data to be processed or already processed by the processor 701 and various modules in the vehicle 700 (e.g., image data, audio data, voice communication data, and video communication data). It can be implemented using flash memory or random access memory (RAM). Data transfer between the processor 701, the communication interface 702, and the memory 703 can be performed via bus 704.

[0139] It should be noted that the descriptions of the storage medium and device embodiments above are similar to the descriptions of the method embodiments above, and have similar beneficial effects. For technical details not disclosed in the storage medium and device embodiments of this application, please refer to the descriptions of the method embodiments of this application for understanding.

[0140] The above embodiments are merely preferred embodiments provided to fully illustrate this application, and the scope of protection of this application is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on this application are all within the scope of protection of this application.

Claims

1. A method for predicting vehicle driving intentions, characterized in that, The prediction method includes: Vehicles within the target distance are selected as graph nodes, and multi-dimensional motion features of each vehicle are extracted to form a node attribute matrix. Traverse all vehicle pairs. If the distance between vehicle pairs is less than a preset distance, construct an edge and calculate the edge weight corresponding to the edge to generate an initial adjacency matrix. Based on the node attribute matrix and the initial adjacency matrix, a multi-stage progressive graph self-enhancement process is performed. At each stage, a graph convolutional network (GCN) with corresponding layers and adaptation parameters are loaded to obtain the processed target vehicle features. The processed target vehicle features are input into the intent classifier to obtain the driving intent output by the intent classifier.

2. The prediction method according to claim 1, characterized in that, Based on the node attribute matrix and the initial adjacency matrix, a multi-stage progressive graph self-enhancement process is performed. At each stage, a graph convolutional network (GCN) of the corresponding number of layers and adaptation parameters are loaded to obtain the processed target vehicle features, including: Based on the node attribute matrix and the initial adjacency matrix, the first stage loads two layers of GCN and the first stage adaptation parameters for processing, and outputs first-order interaction features and car-following recognition results. Based on the first-order interaction features and the updated adjacency matrix obtained in the first stage, the second stage loads 3 layers of GCN and the adaptation parameters of the second stage for processing, and outputs the second-order interaction features and lane change recognition results. Based on the second-order interaction features and the second-updated adjacency matrix obtained in the second stage, the third stage loads a 4-layer GCN and the adaptation parameters of the third stage for processing, and outputs high-order interaction features and full-scene intent recognition results; wherein, the processed target vehicle features include the high-order interaction features.

3. The prediction method according to claim 2, characterized in that, Based on the node attribute matrix and the initial adjacency matrix, the process involves loading a 2-layer GCN and the adaptation parameters of the first stage in the first phase, and outputting first-order interaction features and car-following recognition results, including: Based on the node attribute matrix, two layers of GCN are loaded in the first stage. The multidimensional motion features are mapped to 64-dimensional motion features through the first layer of GCN, and the 64-dimensional motion features are mapped to 32-dimensional motion features through the second layer of GCN. The first-order interaction features are the 32-dimensional motion features. In the first stage, the homogeneity enhancement threshold, homogeneity suppression threshold, enhancement coefficient, and suppression coefficient of the first stage are loaded, and the initial adjacency matrix is ​​dynamically updated to obtain the updated adjacency matrix. In the first stage, the dynamic feature calibration coefficients of the first stage are loaded, and the node features of the first stage are calibrated to the centroids of the two categories of following and non-following to obtain the following recognition result.

4. The prediction method according to claim 3, characterized in that, Based on the first-order interaction features and the updated adjacency matrix obtained in the first stage, the second stage loads a 3-layer GCN and the adaptation parameters of the second stage for processing, outputting second-order interaction features and lane-change recognition results, including: In the second stage, three layers of GCN are loaded. By reusing the weights of the first layer GCN and the second layer GCN, the 32-dimensional motion features are processed to obtain the reconstructed features under the updated adjacency matrix. The reconstructed features are aggregated using a third-layer GCN to obtain 32-dimensional second-order interaction features, wherein the second-order interaction features are the 32-dimensional second-order interaction features. In the second stage, the homogeneity enhancement threshold, homogeneity suppression threshold, enhancement coefficient, and suppression coefficient of the second stage are loaded, and the updated adjacency matrix is ​​dynamically updated to obtain the second-updated adjacency matrix. In the second stage, the dynamic feature calibration coefficients of the second stage are loaded to calibrate the centroids of the second stage node features to three categories: lane keeping, left lane changing, and right lane changing, so as to obtain the lane change recognition result.

5. The prediction method according to claim 4, characterized in that, The adjacency matrix, updated twice based on the second-order interaction features and the second-stage adjacency matrix, is processed in the third stage by loading a 4-layer GCN and the adaptation parameters of the third stage, outputting high-order interaction features and full-scene intent recognition results, including: In the third stage, the homogeneity enhancement threshold, homogeneity suppression threshold, enhancement coefficient, and suppression coefficient of the third stage are loaded, and the adjacency matrix after the second update is dynamically updated to obtain the adjacency matrix after the third update. The adjacency matrix after the three updates is normalized to obtain the normalized adjacency matrix. In the third stage, four layers of GCN are loaded. By reusing the weights of the first layer GCN, the second layer GCN, and the third layer GCN, the 32-dimensional second-order interaction features are processed to obtain the 32-dimensional global adaptation intermediate features under the normalized adjacency matrix. The 32-dimensional global adaptation intermediate features are aggregated by the fourth layer GCN to obtain 32-dimensional full-scene high-order interaction features, wherein the high-order interaction features include the 32-dimensional full-scene high-order interaction features. In the third stage, the dynamic feature calibration coefficients of the third stage are loaded to calibrate the centroids of the third stage node features to six categories: lane keeping, left lane change, right lane change, deceleration, acceleration, and stopping. The calibrated target vehicle features are then input into the classification head to obtain the full-scene intent recognition result.

6. The prediction method according to claim 5, characterized in that, The step of inputting the processed target vehicle features into the intent classifier to obtain the driving intent output by the intent classifier includes: The row vector corresponding to the target vehicle is extracted from the 32-dimensional full-scene high-order interaction features and denoted as the processed target vehicle feature. The processed target vehicle features are input into the intent classifier to obtain the driving intent output by the intent classifier.

7. A vehicle driving intention prediction device, characterized in that, When applied to vehicles, the prediction device includes: The first processing unit is used to select vehicles within the target distance as graph nodes with the target vehicle as the center, and extract multi-dimensional motion features of each vehicle to form a node attribute matrix. The first processing unit is used to traverse all vehicle pairs, construct an edge if the distance between vehicle pairs is less than a preset distance, calculate the edge weight corresponding to the edge, and generate an initial adjacency matrix. The second processing unit is used to perform multi-stage progressive graph self-enhancement processing based on the node attribute matrix and the initial adjacency matrix, loading the corresponding number of layers of graph convolutional network (GCN) and adaptation parameters at each stage to obtain the processed target vehicle features. The prediction unit is used to input the processed target vehicle features into the intent classifier to obtain the driving intent output by the intent classifier.

8. A vehicle, characterized in that, It includes a processor and a memory, the memory storing a computer program that can run on the processor, the processor executing the computer program to implement the method of any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the method described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the method of any one of claims 1 to 6.

Images